[BUG] Azure OpenAI backend: wrong base class, missing MCP/timing/compression support

[ncrispino]

Problem

azure_openai.py extends LLMBackend directly instead of ChatCompletionsBackend. This means it lacks MCP support, custom tool support, StreamingBufferMixin (compression), and duplicates significant code from ChatCompletionsBackend.

Audit against massgen/skills/backend-integrator/SKILL.md found these issues:

Critical

1. Wrong base class — class AzureOpenAIBackend(LLMBackend) should be ChatCompletionsBackend (or at minimum CustomToolAndMCPBackend). Azure OpenAI is OpenAI-compatible, so SKILL.md recommends ChatCompletionsBackend.
   • Claims FilesystemSupport.MCP but has no MCP implementation
   • No custom tool support
   • Duplicates param filtering, chunk conversion, and tool accumulation logic inline
2. Missing timing methods — start_api_call_timing(), record_first_token(), end_api_call_timing() are never called in stream_with_tools(). Breaks observability/metrics.
3. No StreamingBufferMixin — No compression support. _clear_streaming_buffer() and _finalize_streaming_buffer() never called.

Medium

4. Dead workflow tool code — Lines 353-777 contain elaborate workflow tool extraction logic (_build_workflow_tools_system_prompt(), _prepare_messages_with_workflow_tools(), _extract_workflow_tool_calls() with 5 strategies) that is never called from stream_with_tools().

Recommended Fix

Option A (preferred): Refactor to inherit from ChatCompletionsBackend. Remove duplicated code. This gives MCP, custom tools, compression, timing, and workflow tools for free. Azure-specific differences (endpoint config, API version, client creation) can be handled via overrides.

Option B: Keep LLMBackend inheritance but add missing timing calls, remove false FilesystemSupport.MCP claim (return NONE), and either wire up or remove dead workflow tool code.

Files

• massgen/backend/azure_openai.py (~831 lines)
• Tests: massgen/tests/test_azure_openai_backend.py
• Configs: massgen/configs/providers/azure/